4th Slide Set Operating Systems
Prof. Dr. Christian Baun
Frankfurt University of Applied Sciences (1971–2014: Fachhochschule Frankfurt am Main)
Faculty of Computer Science and Engineering
christianbaun@fb2.fra-uas.de
Learning Objectives of this Slide Set
At the end of this slide set You know/understand. . .
the structure, functioning and characteristics of Hard Disk Drives the structure, functioning and characteristics of Solid State Drives the functioning and the most commonly implemented variants of Redundant Array of Independent Disks (RAID)
Exercise sheet 4 repeats the contents of this slide set which are relevant for these learning
objectives
Hard Disk Drives
HDDs are approx. 100 times less expensive per bit versus main memory and they offer approx. 100 times more capacity
Drawback: Accessing data on HDDs is approx. 1000 times slower Reason for the poorer access time:
HDDs are mechanical devices
They contain one or more discs, rotating with 4200, 5400, 7200, 10800, or 15000 revolutions per minute (RPM)
For each side of each disk, an arm exists with a read-and-write head The read-and-write head is used to detect and modify the magnetization of the material
The distance between disk and head is approx. 20 nanometers Also, HDDs have a cache (usually ≤ 32 MB)
This cache buffers read and write operations
Logical Structure of Hard Disk Drives (1/2)
The surfaces of the platters (discs) are magnetized in circular tracks by the heads
All tracks on all disks at a specific arm position are part of a cylinder The tracks are divided into logical units (segments of a circle), which are called blocks or sectors
Typically, a sector contains 512 bytes payload
Sectors are the smallest addressable units of HDDs
Image source (structure): Henry Mühlpfordt. Wikimedia (CC-BY-SA-1.0)
Image source (HDD):purepng.com(CC0)
Logical Structure of Hard Disk Drives (2/2)
If data need be modified, the entire sector must be read and rewritten Today, clusters are addressed by the software (see slide set 6)
Clusters are groups of sectors with a fixed size, e.g. 4 or 8 kB
In file systems of modern operating systems, clusters are the smallest addressable unit of HDDs
Image source (structure): Henry Mühlpfordt. Wikimedia (CC-BY-SA-1.0)
Image source (platter): Tim Bielawa. The Linux Sysadmins Guide to Virtual Disks (CC-BY-SA-4.0)
Addressing Data on Hard Disk Drives
HDDs with a capacity ≤ 8 GB use the Cylinder-Head-Sector addressing CHS faces several limitations:
The Parallel ATA interface and the BIOS offer just. . . 16 bits for the cylinders (up to 65,536)
8 bits for the heads (up to 255. Head number 0 is not used) 8 bits for the sectors/track (up to 255. Sector number 0 is not used)
≤ 7.844 GB can be addressed this way
1024 cylinders * 255 heads * 63 sectors/track * 512 bytes/sector = 8,422,686,720 bytes 8,422,686,720 bytes / 1024 / 1024 / 1024 = 7.844 GB
No 2.5” or 3.5” HDD contains > 16 heads!!!
Logical heads were used
HDDs with a capacity > 7.844 GB use Logical Block Addressing (LBA)
All sectors are numbered consecutively beginning with 0
Logical Block Addressing (LBA)
Image source IT-Handbuch für Fachinformatiker.
Sascha Kersken.
6th edition.
Rheinwerk Verlag
When CHS addressing is used, all tracks contain the same number of sectors
Each sector stores stores 512 bytes of payload
Drawback: Storage capacity is wasted, because the data density decreases from the inner tracks to the outer tracks
When LBA is implemented, this drawback does not exist
Required Time to access Data on HDDs
The access time is an important performance factor 2 factors influence the access time of HDDs
1
Average Seek Time
The time that it takes for the arm to reach a desired track Is for modern HDDs between 5 and 15 ms
2
Average Rotational Latency Time
Delay of the rotational speed, until the required disk sector is located under the head
Depends entirely on the rotational speed of the disks Is for modern HDDs between 2 and 7.1 ms
Average Rot. Lat. Time [ms] =
1000 [ms]
[sec] × 60 [sec]
[min] × 0.5 revolutions
[min]
=
30,000 [ms]
[min]
revolutions [min]
Why does the equation contain 0.5 ?
Once the head has reached the right track, on average a half rotation of the disk must be waited for the correct sector to be under the head =⇒Average Rotational Latency Time = half Rotational Latency Time
Solid State Drives (SSD)
Are sometimes falsely called Solid State Disks Do not contain moving parts
Benefits:
Fast access time Low power consumption No noise generation Mechanical robustness Low weight
The location of data does not matter = ⇒ defragmenting
makes no sense
Image (SSD): Thomas Springer. Wikimedia (CC0)Image (HDD): Erwan Velu.
Wikimedia (CC-BY-SA-1.0)
Drawbacks:
Higher price compared with HDDs of the same capacity Secure delete or overwrite is hard to implement
Limited number of program/erase cycles
Functioning of Flash Memory
Data is stored as electrical charges
In contrast to main memory, no electricity is required to keep the data
Each flash memory cell is a transistor and has 3 connectors Gate = control electrode
Drain = electrode Source = electrode
The floating gate stores electrons (data) Completely surrounded by an insulator Electrical charge remains stable for years
Well written explanation about the functioning of flash memory Benjamin Benz.Die Technik der Flash-Speicherkarten. c’t 23/2006
Reading Data from Flash Memory Cells
A positively doped (p)
semiconductor separates the 2 negatively doped (n) electrodes drain and source
Equal to a npn transistor, the npn passage is not conductive without a base current
Above a certain positive voltage (5V) at the gate (threshold) a n-type channel is created in the p-area
Current can flow between source and drain through this channel If the floating gate contains electrons, the threshold is different
A higher positive voltage at the gate is required, so that current can flow between source and drain
This way the stored value of the flash memory cell is read out
Writing Data into Flash Memory Cells
Data is stored inside flash memory cells by using
Fowler-Nordheim tunneling
A positive voltage (5V) is applied to the control gate As a result, electrons can flow between source and drain
If the high positive voltage is sufficient high (6 to 20V), some electrons are tunneled (= ⇒ Fowler-Nordheim tunneling) through the insulator into the floating gate
This method is also called Channel Hot Electron Injection
Recommended Source
Flash memory.Alex Paikin. 2004.http://www.hitequest.com/Kiss/Flash_terms.htm
Erasing Data in Flash Memory Cells
For erasing a flash memory cell, a negative voltage (-6 to -20V) is applied at the control gate
As a result, electrons are tunneled in the reverse direction from the floating gate
The insulating layer, which surrounds the floating gate, suffers from each erase cycle
At some point the insulating layer is no longer sufficient to hold the charge in the floating gate
For this reason, flash memory survives only a limited number of
program/erase cycles
Functioning of Flash Memory
Memory cells are connected to blocks and (depending on the structure also in) pages
A block always contains a fixed number of pages
Write and erase operations can only be carried out for entire pages or blocks
For this reason, write and erase operations are more complex than read operations
If data in a page need to be modified, the entire block must be erased To do this, the block is copied into a buffer memory
In the buffer memory, the data is modified Next, the block is erased from the flash memory
Finally, the modified block is written into the flash memory
Different Types of Flash Memory
2 types of flash memory exist:
NOR memory NAND memory
The circuit symbol indicates the way, the memory cells are connected
This influences the capacity and access time (latency)
NOR Memory
Each memory cell has its data line Benefit:
Random access for read and write operations
= ⇒ Better latency compared with NAND memory Drawback:
More complex (= ⇒ expensive) construction Higher power consumption than NAND memory Typically small capacities (≤ 32 MB)
Does not contain pages
The memory cells are grouped together to blocks Typical block sizes: 64, 128 or 256 kB
No random access for erase operations
Erase operations can only be done for entire blocks
Fields of application
Industrial environment (e.g. automotive), storing the firmware of a computer system
NOR flash memory (top image) on the IPhone 3G mainboard (bottom image)
Images: Raimond Spekking.
Wikimedia (CC-BY-SA-4.0)
NAND Memory
The memory cells are grouped to pages Typical page size: 512-8192 bytes
Each page has it’s data line
Each block consists of a number of pages Typical block sizes: 32, 64, 128 or 256 pages
Benefit:
Lesser data lines = ⇒ requires < 50% of the surface area of NOR memory Lower manufacturing costs compared with NOR flash memory
Drawback:
No random access = ⇒ Poorer latency compared with NOR memory Read and write operations can only be carried out for entire pages Erase operations can only be carried out for entire blocks
Fields of application
USB flash memory drives,
SSDs, memory cards
Single/Multi/Triple/Quad-Level Cell
4 types of NAND flash memory exist
QLC memory cells store 4 bits TLC memory cells store 3 bits MLC memory cells store 2 bits SLC memory cells store 1 bit SLC storage. . .
is most expensive
provides the best write speed survives most program/erase cycles
SLC memory survives approx. 100,000 - 300,000 program/erase cycles MLC memory survives approx. 10,000 program/erase cycles TLC and QLC memory survives approx. 1,000 program/erase cycles Also memory cells exist, which survive millions of program/erase cycles
Wear Leveling
Wear leveling algorithms evenly distribute write operations File systems, which are designed for flash memory, and therefore minimize write operations, are e.g. JFFS, JFFS2, YAFFS and LogFS
JFFS contains its own wear leveling algorithm
This is often required in embedded systems, where flash memory is
directly connected
Latency of Hard Disk Drives
The performance of CPUs, cache and main memory is growing faster than the data access time (latency ) of HDDs:
HDDs
1973: IBM 3340, 30 MB capacity, 30 ms data access time 1989: Maxtor LXTl00S, 96 MB capacity, 29 ms data access time 1998: IBM DHEA-36481, 6 GB capacity, 16 ms data access time 2006: Maxtor STM320820A, 320 GB capacity, 14 ms data access time 2011: Western Digital WD30EZRSDTL, 3 TB capacity, 8 ms data access time 2018: Seagate BarraCuda Pro ST14000DM001, 14 TB capacity, 4-5 ms data access time
CPUs
1971: Intel 4004, 740 kHz clock speed 1989: Intel 486DX, 25 Mhz clock speed 1997: AMD K6-2, 550 Mhz clock speed
2007: AMD Opteron Santa Rosa F3, 2.8 GHz clock speed 2010: Intel Core i7 980X Extreme (6 Cores), 3.33 Ghz clock speed 2018: AMD Ryzen Threadripper 2990WX (32 Cores), 3 Ghz clock speed 2020: AMD Ryzen Threadripper 3990X (64 Cores), 2.9 Ghz clock speed
The latency of SSDs is ≤ 1 µs = ⇒ ≈ factor 100 better than HDDs But the gap grows because of interface limitations and multiple CPU cores
Further challenge
Storage drives can fail = ⇒ risk of data loss
Enhance latency and reliability of HDDs and SSDs = ⇒ RAID
Redundant Array of independent Disks (RAID)
The performance of the HDDs can not be improved infinitely HDDs contain moving parts
Physical boundaries must be accepted
One way to avoid the given limitations in terms of speed, capacity and reliability, is the parallel use multiple components
A RAID consists of multiple drives (HDDs or SSDs)
For users and their processes, a RAID behaves like a single large drive Data is distributed across the drives of a RAID system
The RAID level specifies how the data is distributed
The most commonly used RAID levels are RAID 0, RAID 1 and RAID 5
Patterson, David A., Garth Gibson, and Randy H. Katz,A Case for Redundant Arrays of Inexpensive Disks (RAID), Vol. 17.
No. 3, ACM (1988)
RAID 0 – Striping – Acceleration without Redundancy
No redundancy
Increases only the data rate
Drives are partitioned into blocks of equal size
If read/write operations are big enough (> 4 or 8 kB), the operations can be carried out in parallel on multiple drives or on all drives
In the event of a drive failure, not the entire data can be reconstructed
Only small files, which are stored entirely on the remaining drives, can be reconstructed (in theory) RAID 0 should only be used when security is
irrelevant or backups are created at regular intervals
RAID 1 – Mirroring
At least 2 drives of the same capacity store identical data
If the drives are of different sizes, RAID 1 provides only the capacity of the smallest drive
Failure of a single drive does not cause any data loss Reason: The remaining drives store the identical data A total loss occurs only in case of the failure of all drives
Any change of data is written on all drives Not a backup replacement
Corrupted file operations or virus attacks take place on all drives
The read performance can be increased by intelligent
distribution of requests to the attached drives
RAID 2 – Bit-Level Striping with Hamming Code Error Correction Each sequential bit is stored on a different drive
Bits, which are powers of 2 (1, 2, 4, 8, 16, etc.) are parity bits
The individual parity bits are distributed over multiple drives
= ⇒ increases the throughput Was used only in mainframes
Is no longer relevant
RAID 3 – Byte-level Striping with Parity Information
Parity information is stored on a dedicated parity drive
Each write operation on the RAID causes write operations on the dedicated parity drive
= ⇒ bottleneck
Was replaced by RAID 5
Payload drives Sum even/odd Parity drive
Bits are 0 + 0 + 0 = ⇒ 0 = ⇒ Sum is even = ⇒ Sum bit 0
Bits are 1 + 0 + 0 = ⇒ 1 = ⇒ Sum is odd = ⇒ Sum bit 1
Bits are 1 + 1 + 0 = ⇒ 2 = ⇒ Sum is even = ⇒ Sum bit 0
Bits are 1 + 1 + 1 = ⇒ 3 = ⇒ Sum is odd = ⇒ Sum bit 1
Bits are 1 + 0 + 1 = ⇒ 2 = ⇒ Sum is even = ⇒ Sum bit 0
Bits are 0 + 1 + 1 = ⇒ 2 = ⇒ Sum is even = ⇒ Sum bit 0
Bits are 0 + 1 + 0 = ⇒ 1 = ⇒ Sum is odd = ⇒ Sum bit 1
Bits are 0 + 0 + 1 = ⇒ 1 = ⇒ Sum is odd = ⇒ Sum bit 1
RAID 4 – Block-level Striping with Parity Information
Parity information is stored at a dedicated parity drive Difference to RAID 3:
Not individual bits or bytes, but blocks (chunks) are stored
P(16-19) = Block 16 XOR Block 17 XOR Block 18 XOR Block 19
Each write operation on the RAID causes write operations on the dedicated parity drive
Drawbacks:
Bottleneck Dedicated parity drive fails more frequently
Seldom implemented, because RAID 5 does not face these drawbacks The company NetApp implements RAID 4 in their NAS servers
e.g. NetApp FAS2020, FAS2050, FAS3040, FAS3140, FAS6080
RAID 5 – Block-level Striping with distributed Parity Information
Payload and parity information are
distributed to all drives Benefits:
High throughput High security level against data loss
No bottleneck
P(16-19) = block 16 XOR block 17 XOR block 18 XOR block 19RAID 6 – Block-level Striping with double distributed Parity Information Functioning is similar to RAID 5
But it can handle the simultaneous failure of up to 2 drives In contrast to RAID 5. . .
is the availability better, but the write performance is lower
is the effort to write the parity information higher
Summary of the RAID Levels
RAID n (number k Allowed Performance Performance
of drives) (net capacity) to fail (read) (write)
0 ≥ 2 n 0 (none) n ∗ X n ∗ X
1 ≥ 2 1 n − 1 drives n ∗ X X
2 ≥ 3 n − [log
2n] 1 drive variable variable
3 ≥ 3 n − 1 1 drive (n − 1) ∗ X (n − 1) ∗ X
4 ≥ 3 n − 1 1 drive (n − 1) ∗ X (n − 1) ∗ X
5 ≥ 3 n − 1 1 drive (n − 1) ∗ X (n − 1) ∗ X
6 ≥ 4 n − 2 2 drives (n − 2) ∗ X (n − 2) ∗ X
X is the performance of a single drive during read or write
The maximum possible performance in theory is often limited by the controller and the computing power of the CPU
If the drives of a RAID 1 have different capacities, the net capacity of a RAID 1 is equal to the capacity of its smallest drive
RAID Combinations
Usually RAID 0, 1 or 5 is used In addition to the popular RAID levels, several RAID
combinations exist At least 2 RAIDs are combined to a bigger RAID
Examples
RAID 00: Multiple RAID 0 are connected to a RAID 0 RAID 01: Multiple RAID 0 are connected to a RAID 1 RAID 05: Multiple RAID 0 are connected to a RAID 5 RAID 10: Multiple RAID 1 are connected to a RAID 0(see figure) RAID 15: Multiple RAID 1 are connected to a RAID 5 RAID 50: Multiple RAID 5 are connected to a RAID 0 RAID 51: Multiple RAID 5 are connected to a RAID 1
Hardware / Host / Software RAID (1/2)
Image Source: AdaptecAdaptec SATA RAID 2410SA
Adaptec SATA II RAID 1220SA
